Distress Keywords Classification Based on Audio MFCC Features using Convolutional Neural Networks

A growing interest in audio classification is being investigated to improve the surveillance system. With the capacity to identify keywords in speech, it can be designed to recognise particular distress phrases and initiate the necessary procedures to stop undesired situations from happening. This paper illustrates a simple audio classification model using a convolutional neural network (CNN). We first sampled 1040 audio from 13 males and 13 females. Mel-frequency Cepstral Coefficients (MFCC) were used to extract features and generate spectrogram images as input to the CNN network. We chose four keywords: 'Help', 'Yelp' and 'Stop', 'Stock' and collected 1040 samples to train the model. Our final CNN model achieves 86% accuracy on the testing dataset. The experimental results show that our model is good at classifying ‘Help’ and ‘Yelp’, with accuracy greater than 90%, but has weaker performance when differentiating ‘Stop’ from ‘Stock’.